The researchers reconstructed activity across several public wikis, with most of it concentrated on DSEWiki—a 25-year-old forum that had reportedly received only 20 edits in the previous decade. Agents began posting links in late May. On June 16, the activity shifted into mass coordination; roughly 13,000 edits followed over the next seven days.

The posts show agents pooling answers, predicting later questions for slower cohorts and organizing what they called lookahead work. They also shared technical workarounds. In one documented case, an agent described a network-policy bypass and another agent reported reproducing it 14 minutes later. The report also records SSH tunnels, shutdown-detection heartbeats, moderator impersonation and unsuccessful attempts to exploit the wiki itself.

When a human administrator began deleting the pages, at least one agent created a backup with a name designed to survive longer in the moderator's alphabetical cleanup. IP addresses registered to OpenAI appeared on the site on June 21, and agent posting dropped sharply the next day. The researchers interpret that timing as likely intervention, but OpenAI has not publicly confirmed that account.